Scientific Computing Seminar

Date:
Tuesday, April 6, 2004
Time:
11:00am-12:00pm
Location:
50F-1647
Seminar Speaker:
Sanjukta Bhowmick
Department of Computer Science and Engineering
The Pennsylvania State University
Title:
Multimethod Solvers: Algorithms, Applications and Software
Abstract:
One of the challenges in scientific computing is the development of efficient computational techniques for the numerical solution of nonlinear partial differential equation (PDE) models. These models arise from a variety of disciplines such as, fluid flow, astrophysics, fusion and others. A popular solution technique for such nonlinear PDEs is Newtons method, which in turn involves the solution of large sparse linear systems. Such systems are typically generated at each Newton iteration and the time taken for their solution can dominate the total simulation time for many applications. A large number of sparse linear solution methods from classes such as direct, preconditioned iterative, multilevel/domain decomposition, etc., are available. However, the performance of these solution methods can vary dramatically depending on factors such as the actual system being solved, across time steps in a time dependent simulation, uniprocessor versus multiprocessor execution, etc. Therefore, it is almost impossible to determine the best sparse linear solver. Our work focuses on developing multimethod solvers using potentially more than one basic solution scheme. I will discuss two classes of multimethod solvers i) adaptive methods that can dynamically select the solution method to match changing numerical properties of the coefficient matrix and, ii)composite methods that use multiple solvers on the same system to provide a more robust solution. I will present our approach for instantiating our multimethod solvers in advanced software environments, including PETSc and the common component architecture framework. I will also report on the serial and parallel performance of our methods on computational fluid dynamics applications such as the driven-cavity flow, the Whitfield code for compressible Euler simulations, and the FUN3d code.
Sponsor of Seminar:
Chao Yang
Scientific Computing

Contact Esmond G. Ng EGNg@lbl.gov